[PATCH v4 6/6] drm/xe: Add last fence as dependency for jobs on user exec queues

[Thomas Hellström]

On 12/6/23 23:21, Matthew Brost wrote:
> The last fence must be added as a dependency for jobs on user exec
> queues as it is possible for the last fence to be a composite software
> fence (unordered, ioctl with zero bb or binds) rather than hardware
> fence (ordered, previous job on queue).
>
> Suggested-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> ---
>   drivers/gpu/drm/xe/xe_exec.c       |  4 ++++
>   drivers/gpu/drm/xe/xe_exec_queue.c |  2 +-
>   drivers/gpu/drm/xe/xe_migrate.c    | 14 +++++++++++---
>   drivers/gpu/drm/xe/xe_sched_job.c  | 17 +++++++++++++++++
>   drivers/gpu/drm/xe/xe_sched_job.h  |  4 ++++
>   5 files changed, 37 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_exec.c b/drivers/gpu/drm/xe/xe_exec.c
> index 438e34585e1e..92b0da6580e8 100644
> --- a/drivers/gpu/drm/xe/xe_exec.c
> +++ b/drivers/gpu/drm/xe/xe_exec.c
> @@ -313,6 +313,10 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   		goto err_put_job;
>   
>   	if (!xe_vm_in_lr_mode(vm)) {
> +		err = xe_sched_job_last_fence_add_dep(job, vm);
> +		if (err)
> +			goto err_put_job;
> +
>   		err = down_read_interruptible(&vm->userptr.notifier_lock);
>   		if (err)
>   			goto err_put_job;
> diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
> index 67e3fd9dfc5f..3911d14522ee 100644
> --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> @@ -887,7 +887,7 @@ static void xe_exec_queue_last_fence_lockdep_assert(struct xe_exec_queue *q,
>   						    struct xe_vm *vm)
>   {
>   	if (q->flags & EXEC_QUEUE_FLAG_VM)
> -		lockdep_assert_held_write(&vm->lock);
> +		lockdep_assert_held(&vm->lock);
>   	else
>   		xe_vm_assert_held(vm);
>   }
> diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
> index e8b567708ac0..ce14498b416a 100644
> --- a/drivers/gpu/drm/xe/xe_migrate.c
> +++ b/drivers/gpu/drm/xe/xe_migrate.c
> @@ -1163,17 +1163,24 @@ xe_migrate_update_pgtables_cpu(struct xe_migrate *m,
>   	return fence;
>   }
>   
> -static bool no_in_syncs(struct xe_sync_entry *syncs, u32 num_syncs)
> +static bool no_in_syncs(struct xe_vm *vm, struct xe_exec_queue *q,
> +			struct xe_sync_entry *syncs, u32 num_syncs)
>   {
> +	struct dma_fence *fence;
>   	int i;
>   
>   	for (i = 0; i < num_syncs; i++) {
> -		struct dma_fence *fence = syncs[i].fence;
> +		fence = syncs[i].fence;
>   
>   		if (fence && !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
>   				       &fence->flags))
>   			return false;
>   	}
> +	if (q) {
> +		fence = xe_exec_queue_last_fence_get(q, vm);
> +		if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
> +			return false;
> +	}
>   
>   	return true;
>   }
> @@ -1234,7 +1241,7 @@ xe_migrate_update_pgtables(struct xe_migrate *m,
>   	u16 pat_index = xe->pat.idx[XE_CACHE_WB];
>   
>   	/* Use the CPU if no in syncs and engine is idle */
> -	if (no_in_syncs(syncs, num_syncs) && xe_exec_queue_is_idle(q_override)) {
> +	if (no_in_syncs(vm, q, syncs, num_syncs) && xe_exec_queue_is_idle(q_override)) {
>   		fence =  xe_migrate_update_pgtables_cpu(m, vm, bo, updates,
>   							num_updates,
>   							first_munmap_rebind,
> @@ -1351,6 +1358,7 @@ xe_migrate_update_pgtables(struct xe_migrate *m,
>   			goto err_job;
>   	}
>   
> +	err = xe_sched_job_last_fence_add_dep(job, vm);
>   	for (i = 0; !err && i < num_syncs; i++)
>   		err = xe_sync_entry_add_deps(&syncs[i], job);
>   
> diff --git a/drivers/gpu/drm/xe/xe_sched_job.c b/drivers/gpu/drm/xe/xe_sched_job.c
> index b467d5bfa4ac..b7d714522ae1 100644
> --- a/drivers/gpu/drm/xe/xe_sched_job.c
> +++ b/drivers/gpu/drm/xe/xe_sched_job.c
> @@ -260,3 +260,20 @@ void xe_sched_job_push(struct xe_sched_job *job)
>   	drm_sched_entity_push_job(&job->drm);
>   	xe_sched_job_put(job);
>   }
> +
> +int xe_sched_job_last_fence_add_dep(struct xe_sched_job *job, struct xe_vm *vm)
> +{
> +	struct dma_fence *fence;
> +
> +	fence = xe_exec_queue_last_fence_get(job->q, vm);
> +
> +	/* Only wait on unsignaled software fences */
> +	if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) &&
> +	    !(fence->context == job->drm.entity->fence_context ||
> +	      fence->context == job->drm.entity->fence_context + 1)) {

Kerneldoc.

Also What's the (fence_context + 1) doing? Otherwise I think the drm 
scheduler code already tests for signaled and same context so we 
shouldn't be needing to code that here as well?

Also we should remember to look at bringing in Rob Clark's work where 
composite fences are unwrapped and added. Not merged yet though AFAICT.

Thanks,

Thomas


> +		dma_fence_get(fence);
> +		return drm_sched_job_add_dependency(&job->drm, fence);
> +	}
> +
> +	return 0;
> +}
> diff --git a/drivers/gpu/drm/xe/xe_sched_job.h b/drivers/gpu/drm/xe/xe_sched_job.h
> index 6ca1d426c036..34f475ba7f50 100644
> --- a/drivers/gpu/drm/xe/xe_sched_job.h
> +++ b/drivers/gpu/drm/xe/xe_sched_job.h
> @@ -8,6 +8,8 @@
>   
>   #include "xe_sched_job_types.h"
>   
> +struct xe_vm;
> +
>   #define XE_SCHED_HANG_LIMIT 1
>   #define XE_SCHED_JOB_TIMEOUT LONG_MAX
>   
> @@ -54,6 +56,8 @@ bool xe_sched_job_completed(struct xe_sched_job *job);
>   void xe_sched_job_arm(struct xe_sched_job *job);
>   void xe_sched_job_push(struct xe_sched_job *job);
>   
> +int xe_sched_job_last_fence_add_dep(struct xe_sched_job *job, struct xe_vm *vm);
> +
>   static inline struct xe_sched_job *
>   to_xe_sched_job(struct drm_sched_job *drm)
>   {